expectedwrong hindsight

Silly Putty Season

AutoGPT and BabyAGI dropped and now the floor is moving.

2 min read 253 words #ai #agents #autogpt #babyagi #2023
hindsight — nailed it

AutoGPT and BabyAGI both faded, exactly as predicted. They didn't work. That wasn't the point. The point was the architecture — an LLM managing its own task list, calling tools, looping — and that architecture became everything. Claude Code, Devin, Cursor agents, CrewAI. The silly putty hardened into the real thing.

There is a specific feeling — you've been keeping up, mostly, skimming the papers, running the demos, generally having a handle on things — and then something drops and you realize you have been standing still while the room was moving.

AutoGPT and BabyAGI are that thing right now.

Not because they work particularly well. They don't, not really — watch one spin for twenty minutes chasing its own tail in a browser window and you'll understand what I mean. But that's not the point. The point is that someone looked at GPT-4 and thought: what if I just... kept asking it what to do next. What if it planned its own tasks. What if the output of one call was the input to the next and you just let it run until something happened.

Nobody designed this. It's penicillin. It's silly putty. It's a graduate student who forgot to clean his petri dish and stumbled into something that shouldn't exist yet.

And now it does exist, and every framework built around the idea of a single well-crafted prompt feels like a horse and buggy.

The uncomfortable thing about these accidental inventions is that they have a way of becoming load-bearing before anyone has time to assess whether they should be. We're going to build a lot of things on top of this. Some of those things will be embarrassing in retrospect. Most of us won't wait to find out which.

I checked the AutoGPT GitHub stars this morning. That was a mistake.